1 // SPDX-License-Identifier: GPL-2.0-or-later 2 /* 3 * Copyright (C) 2010-2017 Mathieu Desnoyers <mathieu.desnoyers@efficios.com> 4 * 5 * membarrier system call 6 */ 7 #include "sched.h" 8 9 /* 10 * For documentation purposes, here are some membarrier ordering 11 * scenarios to keep in mind: 12 * 13 * A) Userspace thread execution after IPI vs membarrier's memory 14 * barrier before sending the IPI 15 * 16 * Userspace variables: 17 * 18 * int x = 0, y = 0; 19 * 20 * The memory barrier at the start of membarrier() on CPU0 is necessary in 21 * order to enforce the guarantee that any writes occurring on CPU0 before 22 * the membarrier() is executed will be visible to any code executing on 23 * CPU1 after the IPI-induced memory barrier: 24 * 25 * CPU0 CPU1 26 * 27 * x = 1 28 * membarrier(): 29 * a: smp_mb() 30 * b: send IPI IPI-induced mb 31 * c: smp_mb() 32 * r2 = y 33 * y = 1 34 * barrier() 35 * r1 = x 36 * 37 * BUG_ON(r1 == 0 && r2 == 0) 38 * 39 * The write to y and load from x by CPU1 are unordered by the hardware, 40 * so it's possible to have "r1 = x" reordered before "y = 1" at any 41 * point after (b). If the memory barrier at (a) is omitted, then "x = 1" 42 * can be reordered after (a) (although not after (c)), so we get r1 == 0 43 * and r2 == 0. This violates the guarantee that membarrier() is 44 * supposed by provide. 45 * 46 * The timing of the memory barrier at (a) has to ensure that it executes 47 * before the IPI-induced memory barrier on CPU1. 48 * 49 * B) Userspace thread execution before IPI vs membarrier's memory 50 * barrier after completing the IPI 51 * 52 * Userspace variables: 53 * 54 * int x = 0, y = 0; 55 * 56 * The memory barrier at the end of membarrier() on CPU0 is necessary in 57 * order to enforce the guarantee that any writes occurring on CPU1 before 58 * the membarrier() is executed will be visible to any code executing on 59 * CPU0 after the membarrier(): 60 * 61 * CPU0 CPU1 62 * 63 * x = 1 64 * barrier() 65 * y = 1 66 * r2 = y 67 * membarrier(): 68 * a: smp_mb() 69 * b: send IPI IPI-induced mb 70 * c: smp_mb() 71 * r1 = x 72 * BUG_ON(r1 == 0 && r2 == 1) 73 * 74 * The writes to x and y are unordered by the hardware, so it's possible to 75 * have "r2 = 1" even though the write to x doesn't execute until (b). If 76 * the memory barrier at (c) is omitted then "r1 = x" can be reordered 77 * before (b) (although not before (a)), so we get "r1 = 0". This violates 78 * the guarantee that membarrier() is supposed to provide. 79 * 80 * The timing of the memory barrier at (c) has to ensure that it executes 81 * after the IPI-induced memory barrier on CPU1. 82 * 83 * C) Scheduling userspace thread -> kthread -> userspace thread vs membarrier 84 * 85 * CPU0 CPU1 86 * 87 * membarrier(): 88 * a: smp_mb() 89 * d: switch to kthread (includes mb) 90 * b: read rq->curr->mm == NULL 91 * e: switch to user (includes mb) 92 * c: smp_mb() 93 * 94 * Using the scenario from (A), we can show that (a) needs to be paired 95 * with (e). Using the scenario from (B), we can show that (c) needs to 96 * be paired with (d). 97 * 98 * D) exit_mm vs membarrier 99 * 100 * Two thread groups are created, A and B. Thread group B is created by 101 * issuing clone from group A with flag CLONE_VM set, but not CLONE_THREAD. 102 * Let's assume we have a single thread within each thread group (Thread A 103 * and Thread B). Thread A runs on CPU0, Thread B runs on CPU1. 104 * 105 * CPU0 CPU1 106 * 107 * membarrier(): 108 * a: smp_mb() 109 * exit_mm(): 110 * d: smp_mb() 111 * e: current->mm = NULL 112 * b: read rq->curr->mm == NULL 113 * c: smp_mb() 114 * 115 * Using scenario (B), we can show that (c) needs to be paired with (d). 116 * 117 * E) kthread_{use,unuse}_mm vs membarrier 118 * 119 * CPU0 CPU1 120 * 121 * membarrier(): 122 * a: smp_mb() 123 * kthread_unuse_mm() 124 * d: smp_mb() 125 * e: current->mm = NULL 126 * b: read rq->curr->mm == NULL 127 * kthread_use_mm() 128 * f: current->mm = mm 129 * g: smp_mb() 130 * c: smp_mb() 131 * 132 * Using the scenario from (A), we can show that (a) needs to be paired 133 * with (g). Using the scenario from (B), we can show that (c) needs to 134 * be paired with (d). 135 */ 136 137 /* 138 * Bitmask made from a "or" of all commands within enum membarrier_cmd, 139 * except MEMBARRIER_CMD_QUERY. 140 */ 141 #ifdef CONFIG_ARCH_HAS_MEMBARRIER_SYNC_CORE 142 #define MEMBARRIER_PRIVATE_EXPEDITED_SYNC_CORE_BITMASK \ 143 (MEMBARRIER_CMD_PRIVATE_EXPEDITED_SYNC_CORE \ 144 | MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED_SYNC_CORE) 145 #else 146 #define MEMBARRIER_PRIVATE_EXPEDITED_SYNC_CORE_BITMASK 0 147 #endif 148 149 #ifdef CONFIG_RSEQ 150 #define MEMBARRIER_PRIVATE_EXPEDITED_RSEQ_BITMASK \ 151 (MEMBARRIER_CMD_PRIVATE_EXPEDITED_RSEQ \ 152 | MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED_RSEQ) 153 #else 154 #define MEMBARRIER_PRIVATE_EXPEDITED_RSEQ_BITMASK 0 155 #endif 156 157 #define MEMBARRIER_CMD_BITMASK \ 158 (MEMBARRIER_CMD_GLOBAL | MEMBARRIER_CMD_GLOBAL_EXPEDITED \ 159 | MEMBARRIER_CMD_REGISTER_GLOBAL_EXPEDITED \ 160 | MEMBARRIER_CMD_PRIVATE_EXPEDITED \ 161 | MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED \ 162 | MEMBARRIER_PRIVATE_EXPEDITED_SYNC_CORE_BITMASK \ 163 | MEMBARRIER_PRIVATE_EXPEDITED_RSEQ_BITMASK) 164 165 static void ipi_mb(void *info) 166 { 167 smp_mb(); /* IPIs should be serializing but paranoid. */ 168 } 169 170 static void ipi_sync_core(void *info) 171 { 172 /* 173 * The smp_mb() in membarrier after all the IPIs is supposed to 174 * ensure that memory on remote CPUs that occur before the IPI 175 * become visible to membarrier()'s caller -- see scenario B in 176 * the big comment at the top of this file. 177 * 178 * A sync_core() would provide this guarantee, but 179 * sync_core_before_usermode() might end up being deferred until 180 * after membarrier()'s smp_mb(). 181 */ 182 smp_mb(); /* IPIs should be serializing but paranoid. */ 183 184 sync_core_before_usermode(); 185 } 186 187 static void ipi_rseq(void *info) 188 { 189 /* 190 * Ensure that all stores done by the calling thread are visible 191 * to the current task before the current task resumes. We could 192 * probably optimize this away on most architectures, but by the 193 * time we've already sent an IPI, the cost of the extra smp_mb() 194 * is negligible. 195 */ 196 smp_mb(); 197 rseq_preempt(current); 198 } 199 200 static void ipi_sync_rq_state(void *info) 201 { 202 struct mm_struct *mm = (struct mm_struct *) info; 203 204 if (current->mm != mm) 205 return; 206 this_cpu_write(runqueues.membarrier_state, 207 atomic_read(&mm->membarrier_state)); 208 /* 209 * Issue a memory barrier after setting 210 * MEMBARRIER_STATE_GLOBAL_EXPEDITED in the current runqueue to 211 * guarantee that no memory access following registration is reordered 212 * before registration. 213 */ 214 smp_mb(); 215 } 216 217 void membarrier_exec_mmap(struct mm_struct *mm) 218 { 219 /* 220 * Issue a memory barrier before clearing membarrier_state to 221 * guarantee that no memory access prior to exec is reordered after 222 * clearing this state. 223 */ 224 smp_mb(); 225 atomic_set(&mm->membarrier_state, 0); 226 /* 227 * Keep the runqueue membarrier_state in sync with this mm 228 * membarrier_state. 229 */ 230 this_cpu_write(runqueues.membarrier_state, 0); 231 } 232 233 void membarrier_update_current_mm(struct mm_struct *next_mm) 234 { 235 struct rq *rq = this_rq(); 236 int membarrier_state = 0; 237 238 if (next_mm) 239 membarrier_state = atomic_read(&next_mm->membarrier_state); 240 if (READ_ONCE(rq->membarrier_state) == membarrier_state) 241 return; 242 WRITE_ONCE(rq->membarrier_state, membarrier_state); 243 } 244 245 static int membarrier_global_expedited(void) 246 { 247 int cpu; 248 cpumask_var_t tmpmask; 249 250 if (num_online_cpus() == 1) 251 return 0; 252 253 /* 254 * Matches memory barriers around rq->curr modification in 255 * scheduler. 256 */ 257 smp_mb(); /* system call entry is not a mb. */ 258 259 if (!zalloc_cpumask_var(&tmpmask, GFP_KERNEL)) 260 return -ENOMEM; 261 262 cpus_read_lock(); 263 rcu_read_lock(); 264 for_each_online_cpu(cpu) { 265 struct task_struct *p; 266 267 /* 268 * Skipping the current CPU is OK even through we can be 269 * migrated at any point. The current CPU, at the point 270 * where we read raw_smp_processor_id(), is ensured to 271 * be in program order with respect to the caller 272 * thread. Therefore, we can skip this CPU from the 273 * iteration. 274 */ 275 if (cpu == raw_smp_processor_id()) 276 continue; 277 278 if (!(READ_ONCE(cpu_rq(cpu)->membarrier_state) & 279 MEMBARRIER_STATE_GLOBAL_EXPEDITED)) 280 continue; 281 282 /* 283 * Skip the CPU if it runs a kernel thread which is not using 284 * a task mm. 285 */ 286 p = rcu_dereference(cpu_rq(cpu)->curr); 287 if (!p->mm) 288 continue; 289 290 __cpumask_set_cpu(cpu, tmpmask); 291 } 292 rcu_read_unlock(); 293 294 preempt_disable(); 295 smp_call_function_many(tmpmask, ipi_mb, NULL, 1); 296 preempt_enable(); 297 298 free_cpumask_var(tmpmask); 299 cpus_read_unlock(); 300 301 /* 302 * Memory barrier on the caller thread _after_ we finished 303 * waiting for the last IPI. Matches memory barriers around 304 * rq->curr modification in scheduler. 305 */ 306 smp_mb(); /* exit from system call is not a mb */ 307 return 0; 308 } 309 310 static int membarrier_private_expedited(int flags, int cpu_id) 311 { 312 cpumask_var_t tmpmask; 313 struct mm_struct *mm = current->mm; 314 smp_call_func_t ipi_func = ipi_mb; 315 316 if (flags == MEMBARRIER_FLAG_SYNC_CORE) { 317 if (!IS_ENABLED(CONFIG_ARCH_HAS_MEMBARRIER_SYNC_CORE)) 318 return -EINVAL; 319 if (!(atomic_read(&mm->membarrier_state) & 320 MEMBARRIER_STATE_PRIVATE_EXPEDITED_SYNC_CORE_READY)) 321 return -EPERM; 322 ipi_func = ipi_sync_core; 323 } else if (flags == MEMBARRIER_FLAG_RSEQ) { 324 if (!IS_ENABLED(CONFIG_RSEQ)) 325 return -EINVAL; 326 if (!(atomic_read(&mm->membarrier_state) & 327 MEMBARRIER_STATE_PRIVATE_EXPEDITED_RSEQ_READY)) 328 return -EPERM; 329 ipi_func = ipi_rseq; 330 } else { 331 WARN_ON_ONCE(flags); 332 if (!(atomic_read(&mm->membarrier_state) & 333 MEMBARRIER_STATE_PRIVATE_EXPEDITED_READY)) 334 return -EPERM; 335 } 336 337 if (flags != MEMBARRIER_FLAG_SYNC_CORE && 338 (atomic_read(&mm->mm_users) == 1 || num_online_cpus() == 1)) 339 return 0; 340 341 /* 342 * Matches memory barriers around rq->curr modification in 343 * scheduler. 344 */ 345 smp_mb(); /* system call entry is not a mb. */ 346 347 if (cpu_id < 0 && !zalloc_cpumask_var(&tmpmask, GFP_KERNEL)) 348 return -ENOMEM; 349 350 cpus_read_lock(); 351 352 if (cpu_id >= 0) { 353 struct task_struct *p; 354 355 if (cpu_id >= nr_cpu_ids || !cpu_online(cpu_id)) 356 goto out; 357 rcu_read_lock(); 358 p = rcu_dereference(cpu_rq(cpu_id)->curr); 359 if (!p || p->mm != mm) { 360 rcu_read_unlock(); 361 goto out; 362 } 363 rcu_read_unlock(); 364 } else { 365 int cpu; 366 367 rcu_read_lock(); 368 for_each_online_cpu(cpu) { 369 struct task_struct *p; 370 371 p = rcu_dereference(cpu_rq(cpu)->curr); 372 if (p && p->mm == mm) 373 __cpumask_set_cpu(cpu, tmpmask); 374 } 375 rcu_read_unlock(); 376 } 377 378 if (cpu_id >= 0) { 379 /* 380 * smp_call_function_single() will call ipi_func() if cpu_id 381 * is the calling CPU. 382 */ 383 smp_call_function_single(cpu_id, ipi_func, NULL, 1); 384 } else { 385 /* 386 * For regular membarrier, we can save a few cycles by 387 * skipping the current cpu -- we're about to do smp_mb() 388 * below, and if we migrate to a different cpu, this cpu 389 * and the new cpu will execute a full barrier in the 390 * scheduler. 391 * 392 * For SYNC_CORE, we do need a barrier on the current cpu -- 393 * otherwise, if we are migrated and replaced by a different 394 * task in the same mm just before, during, or after 395 * membarrier, we will end up with some thread in the mm 396 * running without a core sync. 397 * 398 * For RSEQ, don't rseq_preempt() the caller. User code 399 * is not supposed to issue syscalls at all from inside an 400 * rseq critical section. 401 */ 402 if (flags != MEMBARRIER_FLAG_SYNC_CORE) { 403 preempt_disable(); 404 smp_call_function_many(tmpmask, ipi_func, NULL, true); 405 preempt_enable(); 406 } else { 407 on_each_cpu_mask(tmpmask, ipi_func, NULL, true); 408 } 409 } 410 411 out: 412 if (cpu_id < 0) 413 free_cpumask_var(tmpmask); 414 cpus_read_unlock(); 415 416 /* 417 * Memory barrier on the caller thread _after_ we finished 418 * waiting for the last IPI. Matches memory barriers around 419 * rq->curr modification in scheduler. 420 */ 421 smp_mb(); /* exit from system call is not a mb */ 422 423 return 0; 424 } 425 426 static int sync_runqueues_membarrier_state(struct mm_struct *mm) 427 { 428 int membarrier_state = atomic_read(&mm->membarrier_state); 429 cpumask_var_t tmpmask; 430 int cpu; 431 432 if (atomic_read(&mm->mm_users) == 1 || num_online_cpus() == 1) { 433 this_cpu_write(runqueues.membarrier_state, membarrier_state); 434 435 /* 436 * For single mm user, we can simply issue a memory barrier 437 * after setting MEMBARRIER_STATE_GLOBAL_EXPEDITED in the 438 * mm and in the current runqueue to guarantee that no memory 439 * access following registration is reordered before 440 * registration. 441 */ 442 smp_mb(); 443 return 0; 444 } 445 446 if (!zalloc_cpumask_var(&tmpmask, GFP_KERNEL)) 447 return -ENOMEM; 448 449 /* 450 * For mm with multiple users, we need to ensure all future 451 * scheduler executions will observe @mm's new membarrier 452 * state. 453 */ 454 synchronize_rcu(); 455 456 /* 457 * For each cpu runqueue, if the task's mm match @mm, ensure that all 458 * @mm's membarrier state set bits are also set in the runqueue's 459 * membarrier state. This ensures that a runqueue scheduling 460 * between threads which are users of @mm has its membarrier state 461 * updated. 462 */ 463 cpus_read_lock(); 464 rcu_read_lock(); 465 for_each_online_cpu(cpu) { 466 struct rq *rq = cpu_rq(cpu); 467 struct task_struct *p; 468 469 p = rcu_dereference(rq->curr); 470 if (p && p->mm == mm) 471 __cpumask_set_cpu(cpu, tmpmask); 472 } 473 rcu_read_unlock(); 474 475 on_each_cpu_mask(tmpmask, ipi_sync_rq_state, mm, true); 476 477 free_cpumask_var(tmpmask); 478 cpus_read_unlock(); 479 480 return 0; 481 } 482 483 static int membarrier_register_global_expedited(void) 484 { 485 struct task_struct *p = current; 486 struct mm_struct *mm = p->mm; 487 int ret; 488 489 if (atomic_read(&mm->membarrier_state) & 490 MEMBARRIER_STATE_GLOBAL_EXPEDITED_READY) 491 return 0; 492 atomic_or(MEMBARRIER_STATE_GLOBAL_EXPEDITED, &mm->membarrier_state); 493 ret = sync_runqueues_membarrier_state(mm); 494 if (ret) 495 return ret; 496 atomic_or(MEMBARRIER_STATE_GLOBAL_EXPEDITED_READY, 497 &mm->membarrier_state); 498 499 return 0; 500 } 501 502 static int membarrier_register_private_expedited(int flags) 503 { 504 struct task_struct *p = current; 505 struct mm_struct *mm = p->mm; 506 int ready_state = MEMBARRIER_STATE_PRIVATE_EXPEDITED_READY, 507 set_state = MEMBARRIER_STATE_PRIVATE_EXPEDITED, 508 ret; 509 510 if (flags == MEMBARRIER_FLAG_SYNC_CORE) { 511 if (!IS_ENABLED(CONFIG_ARCH_HAS_MEMBARRIER_SYNC_CORE)) 512 return -EINVAL; 513 ready_state = 514 MEMBARRIER_STATE_PRIVATE_EXPEDITED_SYNC_CORE_READY; 515 } else if (flags == MEMBARRIER_FLAG_RSEQ) { 516 if (!IS_ENABLED(CONFIG_RSEQ)) 517 return -EINVAL; 518 ready_state = 519 MEMBARRIER_STATE_PRIVATE_EXPEDITED_RSEQ_READY; 520 } else { 521 WARN_ON_ONCE(flags); 522 } 523 524 /* 525 * We need to consider threads belonging to different thread 526 * groups, which use the same mm. (CLONE_VM but not 527 * CLONE_THREAD). 528 */ 529 if ((atomic_read(&mm->membarrier_state) & ready_state) == ready_state) 530 return 0; 531 if (flags & MEMBARRIER_FLAG_SYNC_CORE) 532 set_state |= MEMBARRIER_STATE_PRIVATE_EXPEDITED_SYNC_CORE; 533 if (flags & MEMBARRIER_FLAG_RSEQ) 534 set_state |= MEMBARRIER_STATE_PRIVATE_EXPEDITED_RSEQ; 535 atomic_or(set_state, &mm->membarrier_state); 536 ret = sync_runqueues_membarrier_state(mm); 537 if (ret) 538 return ret; 539 atomic_or(ready_state, &mm->membarrier_state); 540 541 return 0; 542 } 543 544 /** 545 * sys_membarrier - issue memory barriers on a set of threads 546 * @cmd: Takes command values defined in enum membarrier_cmd. 547 * @flags: Currently needs to be 0 for all commands other than 548 * MEMBARRIER_CMD_PRIVATE_EXPEDITED_RSEQ: in the latter 549 * case it can be MEMBARRIER_CMD_FLAG_CPU, indicating that @cpu_id 550 * contains the CPU on which to interrupt (= restart) 551 * the RSEQ critical section. 552 * @cpu_id: if @flags == MEMBARRIER_CMD_FLAG_CPU, indicates the cpu on which 553 * RSEQ CS should be interrupted (@cmd must be 554 * MEMBARRIER_CMD_PRIVATE_EXPEDITED_RSEQ). 555 * 556 * If this system call is not implemented, -ENOSYS is returned. If the 557 * command specified does not exist, not available on the running 558 * kernel, or if the command argument is invalid, this system call 559 * returns -EINVAL. For a given command, with flags argument set to 0, 560 * if this system call returns -ENOSYS or -EINVAL, it is guaranteed to 561 * always return the same value until reboot. In addition, it can return 562 * -ENOMEM if there is not enough memory available to perform the system 563 * call. 564 * 565 * All memory accesses performed in program order from each targeted thread 566 * is guaranteed to be ordered with respect to sys_membarrier(). If we use 567 * the semantic "barrier()" to represent a compiler barrier forcing memory 568 * accesses to be performed in program order across the barrier, and 569 * smp_mb() to represent explicit memory barriers forcing full memory 570 * ordering across the barrier, we have the following ordering table for 571 * each pair of barrier(), sys_membarrier() and smp_mb(): 572 * 573 * The pair ordering is detailed as (O: ordered, X: not ordered): 574 * 575 * barrier() smp_mb() sys_membarrier() 576 * barrier() X X O 577 * smp_mb() X O O 578 * sys_membarrier() O O O 579 */ 580 SYSCALL_DEFINE3(membarrier, int, cmd, unsigned int, flags, int, cpu_id) 581 { 582 switch (cmd) { 583 case MEMBARRIER_CMD_PRIVATE_EXPEDITED_RSEQ: 584 if (unlikely(flags && flags != MEMBARRIER_CMD_FLAG_CPU)) 585 return -EINVAL; 586 break; 587 default: 588 if (unlikely(flags)) 589 return -EINVAL; 590 } 591 592 if (!(flags & MEMBARRIER_CMD_FLAG_CPU)) 593 cpu_id = -1; 594 595 switch (cmd) { 596 case MEMBARRIER_CMD_QUERY: 597 { 598 int cmd_mask = MEMBARRIER_CMD_BITMASK; 599 600 if (tick_nohz_full_enabled()) 601 cmd_mask &= ~MEMBARRIER_CMD_GLOBAL; 602 return cmd_mask; 603 } 604 case MEMBARRIER_CMD_GLOBAL: 605 /* MEMBARRIER_CMD_GLOBAL is not compatible with nohz_full. */ 606 if (tick_nohz_full_enabled()) 607 return -EINVAL; 608 if (num_online_cpus() > 1) 609 synchronize_rcu(); 610 return 0; 611 case MEMBARRIER_CMD_GLOBAL_EXPEDITED: 612 return membarrier_global_expedited(); 613 case MEMBARRIER_CMD_REGISTER_GLOBAL_EXPEDITED: 614 return membarrier_register_global_expedited(); 615 case MEMBARRIER_CMD_PRIVATE_EXPEDITED: 616 return membarrier_private_expedited(0, cpu_id); 617 case MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED: 618 return membarrier_register_private_expedited(0); 619 case MEMBARRIER_CMD_PRIVATE_EXPEDITED_SYNC_CORE: 620 return membarrier_private_expedited(MEMBARRIER_FLAG_SYNC_CORE, cpu_id); 621 case MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED_SYNC_CORE: 622 return membarrier_register_private_expedited(MEMBARRIER_FLAG_SYNC_CORE); 623 case MEMBARRIER_CMD_PRIVATE_EXPEDITED_RSEQ: 624 return membarrier_private_expedited(MEMBARRIER_FLAG_RSEQ, cpu_id); 625 case MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED_RSEQ: 626 return membarrier_register_private_expedited(MEMBARRIER_FLAG_RSEQ); 627 default: 628 return -EINVAL; 629 } 630 } 631